wpy/natural-memory-nm21
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
643e22ecb905392eedef8b52ce7c9d0c7f185ed1
natural-memory-nm21/replay_check_v6.md
T
WpyQwq 643e22ecb9 Natural Memory NM2.1: 记忆路由器分叉、数据集缺陷修复与全轴评测证据 
- 引入 MemoryRouterXL 与 v5/v6 流式多线程训练/编码管线
- 修复 prepare_memory_router_dataset 候选池重建缺陷（mega 家族 3568x 加速，输出逐字节相同）
- 修复 v5 被破坏的拒答与多跳标签（train 未知样本 319 -> 16319，multi_hop 平均正例 1.00 -> 2.00）
- 同存储预算下 V2-128 v6 逐轴 22/22 通过：Top-1 41.12% -> 94.62%，未知拒答 0.00% -> 100.00%
- 记录三条被实测推翻的显然优化（logits_to_keep=1 反而慢 55%、XL 容量未带来收益）
- 记忆手术跨架构可移植性 14/14，读写关闭时与原生模型逐位相同
2026-09-19 11:11:31 +08:00

203 lines
9.7 KiB
Markdown
Raw Blame History

| metric | V2-128 v6 final | REPLAY-128 final |
|---|---|---|
| 参数 | 2,037,774 | 2,037,774 |
| router_dim | 128 | 128 |
| 每条记录地址字节 | 512 | 512 |
| **Top-1 正确率** | 94.37% | 94.14% |
| Recall@1 | 88.82% | 88.58% |
| Recall@3 | 96.40% | 96.48% |
| Recall@5 | 97.40% | 97.15% |
| MRR | 95.97% | 95.77% |
| nDCG@3 | 95.31% | 95.37% |
| 多跳证据全中(Top-3) | 95.95% | 96.22% |
| hop 正确率 | 100.00% | 100.00% |
| hop 欠预测率 | 0.00% | 0.00% |
| need F1 | 100.00% | 100.00% |
| need 召回 | 100.00% | 100.00% |
| **未知拒答率** | 100.00% | 100.00% |
| **已知问题被误拒率** | 0.00% | 0.00% |
| 未知问题被误读率 | 0.00% | 0.00% |
| 仲裁准确率 | 100.00% | 100.00% |
| 已知问题 need 概率均值 | 100.00% | 100.00% |
| 未知问题 need 概率均值 | 0.00% | 0.00% |
| 平均分数余量 | 51.3153 | 40.1456 |
| 单查询延迟 ms(GPU) | 0.8989 | 0.9342 |
| 路由 QPS(GPU, 单查询) | 1112.4716 | 1070.4178 |
| 批量 QPS(GPU, batch=64) | 67531.9194 | 70303.4031 |
| 批量 QPS(GPU, batch=256) | 237234.7328 | 244326.0991 |
| 批量延迟 ms(GPU, batch=64) | 0.0148 | 0.0142 |
### family = `benchmark_qa`
| metric | V2-128 v6 final | REPLAY-128 final |
|---|---|---|
| episodes | 64 | 64 |
| Top-1 正确率 | - | - |
| Recall@3 | - | - |
| MRR | - | - |
| hop 正确率 | 100.00% | 100.00% |
| 未知拒答率 | 100.00% | 100.00% |
| 已知问题被误拒率 | - | - |
| 未知问题被误读率 | 0.00% | 0.00% |
| 平均分数余量 | - | - |
### family = `mega_validation`
| metric | V2-128 v6 final | REPLAY-128 final |
|---|---|---|
| episodes | 20000 | 20000 |
| Top-1 正确率 | 98.38% | 98.22% |
| Recall@3 | 99.51% | 99.72% |
| MRR | 99.19% | 99.11% |
| hop 正确率 | 100.00% | 100.00% |
| 未知拒答率 | 100.00% | 100.00% |
| 已知问题被误拒率 | 0.00% | 0.00% |
| 未知问题被误读率 | 0.00% | 0.00% |
| 平均分数余量 | 57.1101 | 44.7334 |
### family = `memory_policy`
| metric | V2-128 v6 final | REPLAY-128 final |
|---|---|---|
| episodes | 1728 | 1728 |
| Top-1 正确率 | 59.40% | 58.76% |
| Recall@3 | 68.42% | 67.31% |
| MRR | 67.60% | 66.33% |
| hop 正确率 | 100.00% | 100.00% |
| 未知拒答率 | 100.00% | 100.00% |
| 已知问题被误拒率 | 0.00% | 0.00% |
| 未知问题被误读率 | 0.00% | 0.00% |
| 平均分数余量 | 0.2715 | -0.2722 |
### family = `native_memory`
| metric | V2-128 v6 final | REPLAY-128 final |
|---|---|---|
| episodes | 128 | 128 |
| Top-1 正确率 | 52.34% | 47.66% |
| Recall@3 | 77.57% | 77.57% |
| MRR | 67.49% | 65.73% |
| hop 正确率 | 100.00% | 100.00% |
| 未知拒答率 | 100.00% | 100.00% |
| 已知问题被误拒率 | 0.00% | 0.00% |
| 未知问题被误读率 | 0.00% | 0.00% |
| 平均分数余量 | -0.8787 | -1.0887 |
### 读取/拒答门槛曲线
| run | threshold | need F1 | need 召回 | 未知拒答率 | 已知问题被误拒率 | 未知被误读率 | tp/tn/fp/fn |
|---|---|---|---|---|---|---|---|
| V2-128 v6 final | 0.30 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
| V2-128 v6 final | 0.40 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
| V2-128 v6 final | 0.50 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
| V2-128 v6 final | 0.60 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
| V2-128 v6 final | 0.70 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
| V2-128 v6 final | 0.80 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
| REPLAY-128 final | 0.30 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
| REPLAY-128 final | 0.40 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
| REPLAY-128 final | 0.50 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
| REPLAY-128 final | 0.60 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
| REPLAY-128 final | 0.70 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
| REPLAY-128 final | 0.80 | 100.00% | 100.00% | 100.00% | 0.00% | 0.00% | 17814/4106/0/0 |
### 按 mega 类别拆解（每类 2,000 条）
| 类别 | episodes | metric | V2-128 v6 final | REPLAY-128 final |
|---|---|---|---|---|
| (未分类) | - | Top-1 正确率 | - | - |
| (未分类) | - | Recall@3 | - | - |
| (未分类) | - | MRR | - | - |
| (未分类) | - | 多跳证据全中 | - | - |
| (未分类) | - | hop 正确率 | - | - |
| (未分类) | - | 未知拒答率 | - | - |
| (未分类) | - | 已知问题被误拒率 | - | - |
| (未分类) | - | 未知问题被误读率 | - | - |
| conflict_update | 2000 | Top-1 正确率 | 99.30% | 99.10% |
| conflict_update | 2000 | Recall@3 | 100.00% | 100.00% |
| conflict_update | 2000 | MRR | 99.65% | 99.55% |
| conflict_update | 2000 | 多跳证据全中 | 100.00% | 100.00% |
| conflict_update | 2000 | hop 正确率 | 100.00% | 100.00% |
| conflict_update | 2000 | 未知拒答率 | - | - |
| conflict_update | 2000 | 已知问题被误拒率 | 0.00% | 0.00% |
| conflict_update | 2000 | 未知问题被误读率 | - | - |
| distractor_128 | 2000 | Top-1 正确率 | 97.50% | 97.30% |
| distractor_128 | 2000 | Recall@3 | 100.00% | 100.00% |
| distractor_128 | 2000 | MRR | 98.75% | 98.65% |
| distractor_128 | 2000 | 多跳证据全中 | 100.00% | 100.00% |
| distractor_128 | 2000 | hop 正确率 | 100.00% | 100.00% |
| distractor_128 | 2000 | 未知拒答率 | - | - |
| distractor_128 | 2000 | 已知问题被误拒率 | 0.00% | 0.00% |
| distractor_128 | 2000 | 未知问题被误读率 | - | - |
| distractor_32 | 2000 | Top-1 正确率 | 98.45% | 98.30% |
| distractor_32 | 2000 | Recall@3 | 100.00% | 100.00% |
| distractor_32 | 2000 | MRR | 99.22% | 99.15% |
| distractor_32 | 2000 | 多跳证据全中 | 100.00% | 100.00% |
| distractor_32 | 2000 | hop 正确率 | 100.00% | 100.00% |
| distractor_32 | 2000 | 未知拒答率 | - | - |
| distractor_32 | 2000 | 已知问题被误拒率 | 0.00% | 0.00% |
| distractor_32 | 2000 | 未知问题被误读率 | - | - |
| forget_correction | 2000 | Top-1 正确率 | - | - |
| forget_correction | 2000 | Recall@3 | - | - |
| forget_correction | 2000 | MRR | - | - |
| forget_correction | 2000 | 多跳证据全中 | - | - |
| forget_correction | 2000 | hop 正确率 | 100.00% | 100.00% |
| forget_correction | 2000 | 未知拒答率 | 100.00% | 100.00% |
| forget_correction | 2000 | 已知问题被误拒率 | - | - |
| forget_correction | 2000 | 未知问题被误读率 | 0.00% | 0.00% |
| long_context | 2000 | Top-1 正确率 | 97.95% | 97.80% |
| long_context | 2000 | Recall@3 | 100.00% | 100.00% |
| long_context | 2000 | MRR | 98.97% | 98.89% |
| long_context | 2000 | 多跳证据全中 | 100.00% | 100.00% |
| long_context | 2000 | hop 正确率 | 100.00% | 100.00% |
| long_context | 2000 | 未知拒答率 | - | - |
| long_context | 2000 | 已知问题被误拒率 | 0.00% | 0.00% |
| long_context | 2000 | 未知问题被误读率 | - | - |
| multi_hop | 2000 | Top-1 正确率 | 98.85% | 99.05% |
| multi_hop | 2000 | Recall@3 | 96.05% | 97.72% |
| multi_hop | 2000 | MRR | 99.42% | 99.52% |
| multi_hop | 2000 | 多跳证据全中 | 92.10% | 95.45% |
| multi_hop | 2000 | hop 正确率 | 100.00% | 100.00% |
| multi_hop | 2000 | 未知拒答率 | - | - |
| multi_hop | 2000 | 已知问题被误拒率 | 0.00% | 0.00% |
| multi_hop | 2000 | 未知问题被误读率 | - | - |
| paraphrase | 2000 | Top-1 正确率 | 99.50% | 99.40% |
| paraphrase | 2000 | Recall@3 | 100.00% | 100.00% |
| paraphrase | 2000 | MRR | 99.75% | 99.70% |
| paraphrase | 2000 | 多跳证据全中 | 100.00% | 100.00% |
| paraphrase | 2000 | hop 正确率 | 100.00% | 100.00% |
| paraphrase | 2000 | 未知拒答率 | - | - |
| paraphrase | 2000 | 已知问题被误拒率 | 0.00% | 0.00% |
| paraphrase | 2000 | 未知问题被误读率 | - | - |
| random_position | 2000 | Top-1 正确率 | 98.05% | 97.80% |
| random_position | 2000 | Recall@3 | 100.00% | 100.00% |
| random_position | 2000 | MRR | 99.02% | 98.89% |
| random_position | 2000 | 多跳证据全中 | 100.00% | 100.00% |
| random_position | 2000 | hop 正确率 | 100.00% | 100.00% |
| random_position | 2000 | 未知拒答率 | - | - |
| random_position | 2000 | 已知问题被误拒率 | 0.00% | 0.00% |
| random_position | 2000 | 未知问题被误读率 | - | - |
| single_fact | 2000 | Top-1 正确率 | 97.45% | 97.05% |
| single_fact | 2000 | Recall@3 | 100.00% | 100.00% |
| single_fact | 2000 | MRR | 98.72% | 98.52% |
| single_fact | 2000 | 多跳证据全中 | 100.00% | 100.00% |
| single_fact | 2000 | hop 正确率 | 100.00% | 100.00% |
| single_fact | 2000 | 未知拒答率 | - | - |
| single_fact | 2000 | 已知问题被误拒率 | 0.00% | 0.00% |
| single_fact | 2000 | 未知问题被误读率 | - | - |
| unknown | 1920 | Top-1 正确率 | 58.99% | 58.10% |
| unknown | 1920 | Recall@3 | 68.96% | 67.92% |
| unknown | 1920 | MRR | 67.60% | 66.30% |
| unknown | 1920 | 多跳证据全中 | 68.96% | 67.92% |
| unknown | 1920 | hop 正确率 | 100.00% | 100.00% |
| unknown | 1920 | 未知拒答率 | 100.00% | 100.00% |
| unknown | 1920 | 已知问题被误拒率 | 0.00% | 0.00% |
| unknown | 1920 | 未知问题被误读率 | 0.00% | 0.00% |
| unknown_abstention | 2000 | Top-1 正确率 | - | - |
| unknown_abstention | 2000 | Recall@3 | - | - |
| unknown_abstention | 2000 | MRR | - | - |
| unknown_abstention | 2000 | 多跳证据全中 | - | - |
| unknown_abstention | 2000 | hop 正确率 | 100.00% | 100.00% |
| unknown_abstention | 2000 | 未知拒答率 | 100.00% | 100.00% |
| unknown_abstention | 2000 | 已知问题被误拒率 | - | - |
| unknown_abstention | 2000 | 未知问题被误读率 | 0.00% | 0.00% |
Reference in New Issue View Git Blame Copy Permalink